A deep learning model can now reconstruct missing portions of 12-lead electrocardiograms well enough to make incomplete clinical records useful for AI diagnostics.
Researchers developed ImputeECG, a Transformer-based autoencoder trained to complete 12-lead, 10-second ECGs from partial recordings. The model was trained on the PTB-XL dataset and tested on PTB-XL, the CPSC2018 dataset, and a real-world cohort of 43,633 records from the Kailuan clinical study. Incompleteness is common in digitized ECG archives — waveforms get cut off by display formats, leads are lost, or signal corruption leaves gaps that render records useless to automated tools. ImputeECG reduced mean absolute error in missing regions by 41.7-51.0% over the strongest baseline on PTB-XL, and by 49.7-51.9% on CPSC2018.
The downstream impact matters more than the reconstruction numbers. In multi-label cardiac classification, ImputeECG restored AUROC to 92.28% on PTB-XL even in the most incomplete scenarios — close to what complete ECGs achieve. On the Kailuan cohort, completing the ECGs improved zero-shot sex prediction AUROC from 82.6% to 85.8% and reduced age prediction error from 10.72 to 9.87 years. That means old, degraded records that would otherwise be discarded can re-enter AI training and diagnostic pipelines.
The real prize here is not a new diagnostic algorithm — it is a data recovery tool. Hospitals are sitting on decades of ECG archives that AI can barely touch because the records are incomplete. ImputeECG is closer to a restoration service than a clinical AI, which is a less glamorous pitch but arguably a more durable one.